| 摘要: | 大腦動靜脈血管畸形(Cerebral Arteriovenous Malformation)為一種先天的血管瘤，其因為動脈與靜脈之間直接相通，缺少微血管因而使靜脈壓力上升，故有破裂的可能。大腦動靜脈血管畸形的大小會隨時間增大，其臨床症狀包含頭痛、頭暈或癲癇等，而破裂則會造成出血性中風進而對患者造成長期甚至永久性的傷害。其處理方式包含保守性觀察、外科手術治療以及放射線治療等。
放射手術(Radiosurgery)為ㄧ種非侵入性治療大腦動靜脈血管畸形的方法。其利用患者神經影像資料，經過電腦模擬運算產生之治療計畫，將高能量放射線束自四面八方照射至計畫治療的區域以提高病灶接受能量，以及降低放射線束對正常腦組織的傷害。然而由於現今依然缺乏對大腦動靜脈血管畸形血管瘤自動化影像分析的方法，因此在治療大腦動靜脈血管畸形的放射手術、影像表現與其臨床反應的研究依然甚少。
本研究回溯分析臺灣一醫學中心自1993年至2014年間，經臨床診斷患有大腦動靜脈血管畸形，並且接受加馬刀放射手術治療之1044位病患的放射手術治療計畫以及腦部核磁共振影像(Magnetic Resonance Imaging)資料。患者的加馬刀放射手術治療計畫中被規劃為血管瘤處的核磁共振影像將被提取，並經由三位神經放射醫師分別獨立標註為病灶(Nidus)、正常腦組織(Brain tissue)以及腦脊髓液(Cerebrospinal Fluid)等三種不同的成分；而本研究利用非監督式機器學習中之K-平均演算法(K-Means)、模糊C聚類(Fuzzy C-means Clustering)以及高斯混合模型(Gaussian Mixture Model)等三種分群法，利用Python語言編程，對上述血管瘤影像進行分群。三種不同非監督式學習模型的預測結果將會與人工標註之結果相互比較，預期能從中獲取最高準確度的模型。
本研究旨在建立大腦動靜脈血管畸形影像自動分群之之人工智慧模型以及不同演算法的效能比較。研究結果顯示使用非監督式機器學習中的模糊C聚類、K-平均演算法以及高斯混合模型皆是潛在的優良分析方法，而高斯混合模型可最有效地分辨出病灶、正常腦組織與腦脊髓液。利用非監督式機器學習的演算法，可協助神經放射醫師在進行放射手術治療規劃時，能方便且快速了解其病灶區域之成分以及現行治療計畫對正常組織造成的傷害，提供精準規劃、精準治療的依據。
Cerebral Arteriovenous Malformation (CAVM) is a congenital cerebral vascular malformation in which arteries and veins are directly connected and lack of capillaries, resulting in the elevation of venous pressure and the potential for rupture. The size of AVM increases over time and its clinical symptoms include headache, dizziness, or epilepsy. When AVM ruptures, it may lead to hemorrhagic stroke and cause long-term or even permanent damage to the patient. Treatment options of the AVM include conservative observation, surgical treatment, and radiation therapy. Radiosurgery is a non-invasive treatment for AVM. It utilizes the patient's neuroimaging data to generate a treatment plan, and then uses high-energy radiation beams from all directions to irradiate the planned treatment area, thus increasing the energy received by the lesion and reducing the damage of the radiation beam to the normal brain tissue. However, due to the lack of automated image analysis methods for AVM, there is still very little research on the radiosurgical treatment, imaging presentation and clinical response of AVM. This study retrospectively analyzed the radiosurgical treatment plans and Magnetic Resonance Imaging (MRI) data of 1044 patients diagnosed with AVM and treated with Gamma Knife radiosurgery from 1993 to 2014 in a medical center in Taiwan. The MRI images of the vascular tumors in the patient's radiosurgical treatment plan were extracted and independently labeled by three neuro-radiologists as nidus, brain tissue and cerebrospinal fluid (CSF), respectively.
This study aimed to establish an artificial intelligence model for the automatic clustering of cerebral arteriovenous malformation (AVM) images, as well as to compare the performance of different algorithms. The results of this research demonstrated that unsupervised machine learning algorithms, including fuzzy C-means clustering, K-means algorithm, and Gaussian mixture models, showed potential as effective analytical methods. Among these, the Gaussian mixture model exhibited the highest performance in distinguishing between nidus, brain tissue, and cerebrospinal fluid. By utilizing unsupervised machine learning algorithms, neuro-radiologists can benefit from a convenient and rapid understanding of the composition of AVM regions and the potential damage to normal tissues caused by current treatment plans during radiotherapy treatment planning. This approach provides a reliable tool for assisting in treatment planning, allowing for more precise and effective treatment strategies, and providing valuable insights into the composition and impact of the lesion and treatment plan. |
